New deploy detected: Lead Scoring with Explainability (v1.2.0), production, Retail Credit BU. Classification: Annex III 5(b) — assessment of creditworthiness. HIGH RISK.
AGENT · AI ACT HIGH-RISK VALIDATOR
Every new AI agent is classified against the AI Act before deploy.
AI Act High-Risk Validator runs the AI Act classification of each new agent or AI system before deploy. It assesses the classification against Annex III (high-risk systems) and Annexes IV/V. Activates the FRIA (Fundamental Rights Impact Assessment) workflow when required.
AI Act High-Risk Validator at work.
FRIA mandatory before deploy. Pre-filled template sent to Carla Young (DPO) and Andrew Ruggeri (AI Officer). Deadline 60 days before go-live.
Received. We'll start the FRIA this week.
Recorded. Deploy blocked in staging until FRIA validation. Event traced in audit.
Why it exists.
The AI Act (Regulation EU 2024/1689) classifies AI systems into four risk categories: minimal, limited, high risk (Annex III), unacceptable. For high-risk systems, the regulation requires specific obligations: registration in the EU registry, FRIA, transparency to the user, human oversight, declared accuracy and robustness.
How it applies the classification.
The operational challenge is not knowing the regulation. It is applying it systematically to every new AI system entering production, without depending on the memory of a single DPO or AI Officer. A missed classification is a regulatory risk; a late classification disrupts delivery timelines. The agent covers the pre-deploy checkpoint. Classification happens in a structured, traceable way for every agent, before it reaches production.
The final decision stays with the AI Act responsible.
The agent identifies ambiguous classification cases and flags them to the AI Act responsible for manual assessment. The final classification stays with the responsible under the customer's procedures. The system does not force a classification when the use case crosses multiple Annex III categories.
Three governance roles, a shared checkpoint.
AI Act responsible
The CDO, AI Officer, or equivalent role gets automatic pre-screening of every new agent. They no longer risk missing the classification of a system that requires FRIA or registration in the EU registry.
Proposal no. 2024-081 In review
Missing disclosure
MiFID II art. · regulated financial instrument
Alt. 1 …in compliance with MiFID II and applicable supervisory provisions.
Alt. 2 …with full disclosure attached to the offer document.
Audit trace recorded · 14:31
DPO
Has an inspectable trace of AI Act classifications for audits by the data protection authority or the national AI Act authority. Queries on the registry return the full classification chain, FRIA and deploy authorisation for every system.
fnol.receive 09:14:22 ALLOW
triage.classify 09:14:25 ALLOW
idd.check 09:14:31 WARN
liquidation.propose 09:15:02 ALLOW
SELECT * FROM audit_log WHERE claim_id = '2024-0847'
Compliance lead
Sees the agent inventory aggregated by risk category. The runtime's regulatory exposure is measurable and manageable. Typical relevance in banking, insurance, healthcare, public sector, energy.
KYC case KYC-2024-091 In verification
ID document VERIFIED
Biometrics · SCA PSD2 LINKING OK
Beneficial owner BUSINESS REGISTRY
Case forwarded to AML Screening
A new scoring agent is classified and blocked in staging until the FRIA is complete.
The deploy triggers the agent automatically.
A bank is about to deploy a new Lead Scoring with Explainability agent for consumer credit. The deploy triggers AI Act High-Risk Validator automatically. The agent reads the description of the new system: use case 'preliminary assessment of customers for consumer credit', data 'personal details + income declarations + public data', output 'scoring + reason code'. It compares against Annex III.
Annex III 5(b): HIGH RISK. FRIA activated.
The system falls within Annex III point 5(b) — systems intended to assess the creditworthiness of natural persons. Category: HIGH RISK. The agent activates the FRIA workflow: notification to the DPO and the AI Act responsible with the system description, the applied classification, an explicit reference to Annex III 5(b), a pre-filled FRIA template, and the deadline for completion.
DPO and AI Officer validate and authorise the deploy.
The DPO and the AI Act responsible validate the classification, complete the FRIA, and authorise the deploy. The event — classification, FRIA activated, decisions, authorisation — stays in the runtime audit registry, queryable with a standard SQL client.
Configuration and technical resources.
The rules are declarative. The customer's AI Act responsible, DPO and compliance team define in a readable format the classification rules (use-case to AI Act Annex mapping, assessment criteria, FRIA template), the automatic block thresholds (systems classified as "unacceptable risk"), the notification and escalation procedures. The rules live in the customer's repository, versioned.
The AI Act comes into force gradually: different provisions have different application dates. The classification rules repository requires periodic updates by the customer's compliance team. The agent runs the configured rules; the responsibility for the completeness of the mapping against the updated regulatory picture rests with the compliance team.
- Language
- TypeScript (Node.js)
- LLM model
- customer's choice: Anthropic, OpenAI, Mistral, open source models hosted internally, AWS Bedrock for a private model
- Built-in controls used
- pii-detector, prompt-injection, topic-guardrail
- Trigger
- new agent deploy in the Polyant runtime + import of external AI systems into the customer's AI Act perimeter
- AI Act rules
- declarative, versioned, written by the AI Act team + DPO; periodic updates by the compliance team per the regulation's application calendar
- FRIA templates
- declarative, versioned
- Classification note
- completeness of the mapping against the gradually applicable AI Act provisions is the responsibility of the customer's compliance team
- Memory
- persistent per instance
- Registry
- immutable, AI Act audit inspectable with a standard SQL client
Frequently asked questions about the agent.
For agents classified as 'unacceptable risk' (prohibited by the AI Act, e.g. mass social scoring), yes: automatic block with notification to the AI Act responsible. For agents classified as 'Annex III high-risk', the agent activates the FRIA workflow but does not block the deploy on its own: the block is managed by the customer's procedure pending FRIA completion and authorisation from the responsible parties.
The agent identifies ambiguous classification cases and flags them to the AI Act responsible for manual assessment. The final classification stays with the responsible under the customer's procedures. The system does not force a classification when the use case crosses multiple Annex III categories.
The AI Act has staggered application dates for different provisions (some already in force, others from 2025-2026). The customer's compliance team updates the repository each time new provisions come into effect. The recommended pattern is a quarterly review with the AI Act responsible to verify the consistency of the mapping.
The typical pattern is 12-18 weeks. Discovery and mapping of the customer's AI Act perimeter 2-4 weeks, writing the declarative rules with the compliance team 6-8 weeks, integration with the existing FRIA workflow 2-3 weeks, hand-off 1-2 weeks.
From a 30-minute conversation to the squad in production.
A 30-45 minute conversation to understand how AI Act High-Risk Validator would configure to the customer's case. Which AI system perimeter, which classification rules, which FRIA workflow already in use.